Michael Kuron authored 4 years ago and Markus Holzer committed 4 years ago

Some optional components use Boost MultiArray and PropertyTree, so we still include the headers if available.
We also still search std::experimental for any, filesystem and optional because some otherwise C++17-compliant compilers did not support them in std until recently.

Markus Holzer authored 4 years ago and Dominik Thoennes committed 4 years ago

also clean up some old leftovers nearby

It is not needed because deprecations are compiler warnings and should not be ignored anyway.
Also, no need to export WARNING_DEPRECATED to the C++ code.

We need to have at least one other required library, otherwise FindBoost will consider itself successful even if it didn't find anything.
Also, add Python 3.8 support by using the Python version numbers as library version numbers.

use cuda as a first level language in CMake with `enable_language(CUDA)`
enable tests where cuda and codegen are both activated
various minor fixes

CMake: don't use CMAKE_CXX14_STANDARD_COMPILE_OPTION directly, instead go through CMAKE_CXX_STANDARD

markl

This option is OFF by default.
If turned on cmake will notify if a target cannot be build due to a missing module.

- workaround seems to be not necessary any more, for recent IntelMPI versions
- the workaround did not play with CMake's try_compile function, which tries to compile some test programs to check for compiler features
  The workaround added '-include mpi.h' to the compile flags, which are also used by try_compile, but try_compile does not set include directories with "-I"
  This lead to the problem, that for example the OpenMP flag was not correctly determined, since there the try_compile mechanism is used

- recent GCC version produce error when "-isystem /usr/include" is added
  to command line. Error is: '#include_next <math.h>' not found
- the removed "hack" added the cuda include dir with "-isystem" to
  the CUDA compiler arguments to prevent warnings from boost header
  -> currently there seem to be no warnings generated from boost headers
     so "hack" is removed

added filter to exclude external libraries

- NearestNeighborFieldInterpolator::get needs to break out of all loops, just not the current one. While the test seemed to succeed anyway, it crashed with corrupted memory during exit when compiled with Intel 17 in Release mode.
- FftTest should not use thread-unsafe RNG in parallel section. Otherwise the Intel 16 compiler crashes.
- Revert to -openmp on Intel < 16.0.3, which crashes while processing some of our OpenMP loops (e.g. BoundaryHandling::treatDirection).

Also adds OpenMP support for macOS (requires Xcode 7.0, CMake 3.12 and separate libomp from MacPorts/Homebrew)